OPTIMIZATION OF NEURAL NETWORK ARCHITECTURE FOR ... · OPTIMIZATION OF NEURAL NETWORK ARCHITECTURE...

International Journal of Artificial Intelligence and Applications (IJAIA), Vol. 7, No. 5, September 2016

DOI: 10.5121/ijaia.2016.7501 1

OPTIMIZATION OF NEURAL NETWORK

ARCHITECTURE FOR BIOMECHANIC

CLASSIFICATION TASKS WITH

ELECTROMYOGRAM INPUTS

Alayna Kennedy1

and Rory Lewis2

1 The Pennsylvania State University, University Park, PA, USA

2 Department of Computer Science, University of Colorado at Colorado Springs, USA

ABSTRACT

Electromyogram signals (EMGs) contain valuable information that can be used in man-machine

interfacing between human users and myoelectric prosthetic devices. However, EMG signals are

complicated and prove difficult to analyze due to physiological noise and other issues. Computational

intelligence and machine learning techniques, such as artificial neural networks (ANNs), serve as powerful

tools for analyzing EMG signals and creating optimal myoelectric control schemes for prostheses. This

research examines the performance of four different neural network architectures (feedforward, recurrent,

counter propagation, and self organizing map) that were tasked with classifying walking speed when given

EMG inputs from 14 different leg muscles. Experiments conducted on the data set suggest that self

organizing map neural networks are capable of classifying walking speed with greater than 99% accuracy.

KEYWORDS

Electromyogram, Prostheses, Neural Networks, Biomechanical Analysis, Machine Learning, Myoelectric

Control Schemes, Self Organizing Maps, Pattern Recognition Algorithms

1. INTRODUCTION

Assistive prosthetic devices commonly use a myoelectric control scheme, which adjusts the

function of a prosthetic device given electromyogram (EMG) inputs, and which has been proven

an effective method of man-machine interfacing between user and prosthetic [1]-[4]. However,

despite the significant development of the prosthetic industry over the past decade, high-accuracy

commercial prostheses remain too expensive for the average middle-class amputee to afford [5],

[6]. Even the most accurate devices possess multiple issues, including difficulty actuating

multiple degrees of freedom and low accuracy without a high number of electrodes [7]. Within

academia, computational intelligence pattern recognition techniques of EMG analysis provide

accurate results, but are often computationally expensive [8]. Therefore, academic models result

in limited real-world improvements in the control of prosthetic devices, since available cheap


2

prostheses do not possess the computation power necessary to run complex pattern recognition

algorithms. An ideal prosthetic control scheme would be able to achieve accuracy in classification

tasks while remaining simple enough to be implemented in inexpensive devices [6], [8].

Among the common electrophysiological signals, EMG recordings are used most extensively in

man-machine interfacing because of their non-invasiveness, relatively easy application, and

richness of neural information [3], [8]. While other biosignals like electroencephalograms (EEGs)

and electrooculograms (EOGs) can be used to predict human movement, EMG signals have

become the standard biosignal for myoelectric prostheses because they directly transmit electrical

signals from the muscle during periods of contraction or relaxation [3], [9], [10]. In addition,

EMG signals have been shown to precede muscle kinematics by 100ms; therefore, they can be

used to predict human movements to create a prosthetic device that would analyze EMG signals

and transmit them to a prosthetic, which would then move as if it were the amputee’s biological

limb [1], [2].

The EMG signal is the sum of the electrical activity of the muscle fibers, as triggered by the

impulses of activation of the innervating motor neurons [3], [8], [10]. Surface EMGs are obtained

by convolution of each motor neuron spike train by the motor unit action potential and have

mathematically the same expression as the neural efferent signal [10].

Equation 1 shows a simple model of the EMG signal:

Equation 1

where x(n) is the modeled EMG signal, e(n) represents the firing impulse, h(r) represents the

motor unit action potential, w(n) is the zero mean additive white Gaussian noise, and N is the

number of motor unit firings [8].

Despite the rather direct relation between a motion and the expressed EMG signal, however, there

remain several open issues before the control of prostheses by EMG will reach the ideal

characteristics needed for widespread acceptance by patients [4], [8]. Raw EMG data alone

provides very valuable information about human biomechanics in a fairly useless form, since the

amount of noise in the data require it to be processed, quantified, cleared of noise, and

decomposed into individuals MUAPs (Fig 1.) before it can be used in any significant way [7],

[10], [11]. While difficult, the effective analysis of EMG signals can be achieved through wavelet

analysis, artificial neural networks (ANNs), and different clustering algorithms [10].


3

Fig. 1: Decomposition of EMG signals into individual MUAPs

While many studies have been done on the continuous analysis of EMG data using neural

networks and complex pattern recognition algorithms, little work has been done on less

computationally expensive classification tasks [4]. While regression tasks provide detailed

information in the form of bodily spatial coordinates, they also require more electrodes and

achieve a lower degree of accuracy than many classification tasks [7]. Although classification

tasks only output a discrete category, they can also provide a great deal of information about

future movement. For example, walking speed can determine multiple factors about human gait

[12]. Other applications of classification tasks to prosthetic technology can be found in Gandolla

and Gehani's studies, where trained ANNs classified distinct hand movements, resulting in a hand

prosthetic performing the movement [1], [3].

This study aims to find a network architecture that can accurately classify kinematic information

from EMG signals while remaining simple enough for application in affordable prostheses, since

one of the most significant applications of EMG analysis technology is within the field of

orthotics and prosthetics [5].

This study aims to find the best way to use ANNs to classify human movement from EMG data,

by measuring the performance of multiple neural network architectures when tasked with

classification tasks. The different ANN architectures will attempt to classify walking speed into

one of five categories given surface EMG inputs from 14 different leg muscles [13].

2. PREVIOUS WORK

The parallel computation power and nonlinear operations performed by the human brain inspired

the original ANNs, so when creating biomimetic control schemes for prosthetics, many

researchers turned to these networks to create efficient myoelectric devices that worked with the

human body [2]. While some of the earliest attempts to create myoelectric prostheses date back to

the 1970s, advanced pattern recognition techniques did not emerge until the 1990s with the rise of

more accurate machine learning techniques like neural network backpropagation and time-series

analysis of inputs [3], [14]. One of these developments in myoelectric control was a dynamic


4

recurrent neural network which accurately predicted spatial coordinates of arm trajectory [15].

The development of recurrent networks, in which all neurons are interconnected, resulted in a

marked improvement in the accuracy of spatial coordinate prediction of the body given EMG

inputs. Networks like Draye’s and Cheron’s also implemented time-series backpropagation,

which took into account the previous values of EMG input instead of just the immediate input

channel [14], [15].

Other networks modeled EMG inputs with an auto-regressive (AR) model, and then passed them

through an ANN to control the movements of a virtual prosthetic [3]. The results from these

studies have shown that controlling prostheses through an ANN recognition of EMG patterns can

optimize the number of electrodes, provide greater degrees of freedom for the device, and

accurately predict user intent before muscle movements [1], [7], [20]. Two important network

architectures for classification of EMG data are the Kohonen Network, and a cascade architecture

network with a preprocessing step involving Kohonen maps.

2.1. Kohonen Map

Also referred to as a self-organizing map (SOM), this network implements unsupervised learning,

which maps points in the input space to points in the output space while preserving the topology

[16]. Normally, the input space is of high dimension while the output is usually two dimensional.

The network identifies the spatial concentration of the network activity that is best tuned to the

present input [3], [16].

Fig. 2: Detailed view of a Kohonen self organizing map network

The following six steps explain the algorithm for producing Kohonen maps:


5

Step one: Select network topology: The arrangement of the clusters for the network can be

square, circular, etc.

Step Two: Initialize weights to random values: The weight matrix represents the connections

between the neurons of the network, and it is randomized to initial values which are subsequently

modified during the learning phase

Step Three: Select a pattern: Chooses an input pattern, “x”, from the input examples.

Step Four: Find best matching unit: The node with the weight vector most similar to the input

vector, defined as the node with the smallest Euclidean distance to the input weights, is selected

as the matching node.

Step Five: Update weights to all nodes: The “winning” node and its topological neighborhood

are updated by the SOM algorithm according to the equation:

Equation 2

Where mi(t+1) is the new weight, and mi(t) is the old weight, α(t) is the learning rate factor

(0<α(t)<1) and t = 0,1,2... is an integer representing the discrete time coordinate.

Step Six: Iteration: Repeat steps 1 to 5 for all input patterns and the repeat for a pre-determined

number of iterations [3], [16].

SOMs have been used widely in many applications, including in EMG clustering classification.

This network can be used as a preprocessing stage for other ANN architectures such as a Cascade

Architecture with Feature Maps (CANFM).

2.2. Cascade Architecture with Feature Maps (CANFMs):

The CANFM network implements a cascaded architecture of neural networks with feature maps

(CANFM) [17]. This network first passes the data through an unsupervised Kohonen self-

organizing map, outputting 2D coordinates onto the x and y axes of the 2D topological net, which

both reduces the input dimensions of the EMG data channels and removes some noisy data from

the original inputs [16], [17].

This network first randomizes the initial weights of the network, then passes the values through a

self-organizing map (SOM) where the EMG values are clustered. The unsupervised SOM can

find a winning neuron on the 2-D topology map to represent the original pattern, and the x and y

coordinate of this winning neuron y\textsubscript{c} become the input values for a back-

propagation neural network (BPNN) [17], [18]. After reduction of the input space using

Kohonen's SOM, the three sets of 2-D coordinates (six newly condensed features) are fed into the


6

BPNN for further classification. Huang chose the BPNN as the post-classifier of CANFM

because of its learning ability and fast recall speed. In the case of Huang's study, there are eight

postures to be classified, so the BPNN has eight output nodes (Fig. 3) [17].

Fig. 3: Detailed information of CANFM

3. ACQUISITION OF EMG DATA

The EMG data for this study was collected by Hof et. al and made available to the public via the

Clinical Gait Analysis Database [13].

3.1 Collecting Data

Surface EMGs of 14 leg muscles were recorded from two homogeneous groups (n=9 and 11,

respectively) of young healthy male subjects (mean age 22 years (S.D. 1.5), stature 1.85 m (S.D.

0.05), leg length 0.98 m (S.D. 0.04), body mass 73 kg S.D. 8). The average personal data of both

groups was matched to compensate for the division in the two groups. Subjects walked barefoot

on a 10 m indoor walkway at speeds of 0.75, 1.00, 1.25, 1.50, and 1.75 m s-1

.


7

3.2 Pre-processing: Filtering and Blocking

Compared with other biosignals, EMG signals are difficult to analyze due to their small

amplitude, which makes them highly subject to both internal and external electrical noise.

Internal noise occurs when an electrode picks up signals from more than one MUAP and overlaps

their signals, while external noise can result from equipment noise, electromagnetic radiation, or

motion artefacts. Therefore, pre-processing and filtering is essential to obtain reliable raw EMG

data [3], [10], [13].

In this study, the EMGs were high-pass filtered at 20 Hz, rectified and smoothed with a 25 Hz

third order Butterworth low-pass filter. Smoothed rectified EMGs were, after A/D conversion

with a sample frequency of 100 Hz, linearly interpolated to 100 points per stride, triggered by

heel contact of the leg of interest. The recorded steps were screened to exclude those with

obvious artefacts or incorrect foot contacts. In this way for every individual i, normalized speed v̂,

and muscle m, average individual profiles e(p, m, v, i) were determined from at least 10 steps

over p=1-100% of the gait cycle [13].

Fig. 4: Average EMG profiles of vastus medialis muscle after normalizing. Also shows low and high limits.

Time is given as a percentage of stride, starting at heel contact, but in a scale running 0-100-50%, in order

to represent better the activity around heel strike.

4. NEURAL NETWORK CREATION

The EMG data taken from the Hof study was split into training, validation, and testing data in a 2-

1-1 ratio, as is considered best practice when training neural networks [19]. The data was then

normalized to a range of values between 0 and 1 using the following equation [12].


8

Equation 3

where x is the value to be normalized, d represents the high and low values of the data, and n

represents the high and low normalization range desired [20].

Four different neural networks architectures were created, a multilayer feedforward perceptron

network, a recurrent neural network, a self organizing map (SOM), and a counter propagation

neural network (CPN). Each network performed classification tasks on the EMG data, and their

performance was recorded. In the first type of architecture, feedforward perceptrons, the networks

all had a single hidden layer and were tested with four different combinations of activation

functions: linear, radial basis function (RBF), sigmoid (SIG), and hyperbolic tangent function

(TANH) [21]. The TANH activation function uses the hyperbolic tangent function

Equation 4

and has proved a powerful tool for classification tasks [20]. However, in the analysis of EMG

data with neural networks, the SIG function

Equation 5

is more typically used [15], [19]. The feedforward networks were also tested with varying

numbers of hidden layer neurons to ascertain the optimal number for both speed and accuracy of

classification. All of the feedforward networks implemented a resilient backpropagation

(RPROP) learning algorithm, which is a gradient-based optimization technique similar to the

more common regular backpropagation. RPROP is often faster than training with

backpropagation and does not require any free parameter values to be specified. RPROP works

similarly to traditional backpropagation, except an individual delta value is calculated for each

connection. These delta values are gradually changed until the neural network weight matrix

converges on a potentially ideal weight matrix [19], [20], [22].

In each iteration of RPROP, the new weights are given by [22]:


9

Equation 6

The second type of architecture tested was a recurrent neural network (RNN) architecture. RNNs

are a subset of ANNs where connections between units form a directed cycle. This allows the

internal state of the network to exhibit dynamic temporal behavior, using their internal memory to

process sequences of inputs. The two types of RNNs tested were a three-layer Elman network and

a Jordan network. In the Elman simple recurrent network, context units connected to the hidden

layer maintain a copy of the previous values of the hidden units, allowing the network to maintain

a memory of the previous time step [23]. Jordan networks are similar, but the context units are

fed from the output layer instead of the hidden layer [24].

Fig. 5: A structure of the trained Elman-Jordan neural network

The third type of architecture tested were SOM neural networks. This network, unlike the

previous two, was trained using unsupervised learning to produce a two-dimensional discretized

representation of the input space of the training samples, called a topology map. In addition, the

SOM applied competitive learning techniques to improve the network. The SOM architecture was

tested with two different training algorithms: K-means nearest neighbor training and cluster copy

training.

The fourth network type, counter propagation networks (CPNs) are similar to CANFM networks,

consisting of an outstar network and a competitive filter network. Each neuron in the input layer

is processed through a Kohonen network which categorizes the input pattern, serving as the

hidden layer for the network [25]. The outputs of the Kohonen map are then filtered through an

outstar array which reproduces the correct output pattern for the category. Training is done in two


10

stages; first the hidden layer learns to categorize the patterns and then the weights for that layer

become fixed. Then the output layer is trained. One of the advantages of CPN and Kohonen

networks is that the training phase requires a relatively small number of epochs, usually several

hundred, which is considered a tiny number compared to other ANNs, such as convolutional and

deep neural networks which require thousands of iterations of training [3].

For each of the four network architectures chosen, all of the networks were trained, tested, and

evaluated for error four times, then the error results over each of the four runs were averaged.

Each network was trained using the training and validation, then the method was evaluated using

the testing data. When training the network, a maximum error percentage of 1%, max step of 50

and initial update of 0.1 were used [26].

5. RESULTS FROM THE COMPARISON OF NETWORK ARCHITECTURES

Initial results on optimal classification for multiple neural network architectures corroborated

results from previous studies, showing that in feedforward networks, the number of neurons in the

hidden layer does not significantly affect accuracy of classification, but a higher number of

hidden neurons allows the network to train faster. The feedforward networks tested all had 14

inputs, corresponding to the 14 EMG channels, and used a TANH function for both the hidden

and output layer activation.

First, the network was tested with 7 hidden layer neurons, half the number of input neurons,

which resulted in a 4.782% error in 102 seconds of training time. The number of hidden layers

was increased by two during each subsequent experiment, and both the error and training time

recorded, with the final test consisting of a network with 30 hidden layer neurons. While error

decreased slightly with a greater number of hidden neurons, the more significant change was the

decrease in training time with more hidden layer nodes. Both the error and the training time

leveled off at around 28 hidden layer neurons, with an error value of 4.4% and a training time of

0.35 seconds. Therefore, we can state generally that the optimal number of hidden neurons to

minimize both training time and error for this experiment is about twice the number of input

neurons, a result which has been produced in several previous studies [25].


11

Fig. 6: Graph of training time and percent error of feedforward network vs the number of hidden neurons in

the hidden layer of the perceptron.

In analyzing the performance of the feedforward neural network architecture, we can see that

using the TANH activation function for both the hidden layer and the output layer activation

provides the most accurate results for classification tasks, with an average error of 0.99%. Using

the SIG activation function in the hidden layer and output layer activation resulted in an error of

12%, a result which is surprising since most neural networks with EMG inputs use the sigmoid

activation function with relative success. However, most studies using the SIG function utilize

networks which perform regression tasks to output spatial coordinates, while our network

classified the EMG data into 5 discrete walking speeds.

Furthermore, both the feedforward network with TANH activation for the hidden layer and linear

outputs and the feedforward network with an RBF activation function performed poorly when

classifying EMG data, with average errors of 15%. While the TANH & linear network was

included primarily to serve as a control for the TANH & TANH network, the RBF function has

been shown to perform well in networks that classify numerical data. However, these results

suggest that RBF networks do not perform well on classification tasks with small-amplitude

EMG data.


12

Fig. 7: Average error percentages of four feedforward networks with varying activation functions and two

SOM with different training techniques.

The performance of recurrent neural networks on classification tasks was comparable to the

TANH and linear feedforward neural network, with the Jordan recurrent network producing an

error of 15.3% and the Elman recurrent network producing an error of 13.7%. While dynamic

recurrent neural networks have previously performed accurate classifications of EMG data when

diagnosing neuromuscular diseases or prehensile human postures, they do not appear to perform

well when given normalized EMG inputs to classify walking speed [17], [24]. This result could

be due to the generalized normalization of the data to values between 0 and 1, which has been

shown to improve results in feedforward networks, but is often not used with recurrent systems

[23].

Fig. 8: Average error percentages of four feedforward networks with varying activation functions and two

SOM with different training techniques.


13

Although the TANH feedforward network classified walking speed with an error of less than 1%,

the most effective networks for classifying EMG data were SOM neural networks. Both SOM

networks had an average error rate of under 2.5\%, and the SOM network trained with a cluster

copy training algorithm had an error rate of 0.27\%. These results give the cluster copy trained

SOM in this study an average accuracy percentage of 99.73%, a result which improves upon

many machine learning techniques of EMG classification.

Fig. 9: Comparison of the accuracy percentages obtained from multiple machine learning techniques,

including the cluster copy training SOM created in this study.

Table 2 compares the results obtained in this study from the cluster copy training (CCT) SOM

with results from different machine learning techniques of EMG classification [6][17][27]. Most

of the studies cited in this table classified gestures or movements, not walking speed. For

example, the adaptive neuro-fuzzy inference system (ANFIS), classified input EMG data into one

of four different hand movements [6]. Although the classification tasks differ slightly from those

in this study, the techniques are similar enough for a comparison to be made between the results.

Each study utilized various machine learning methods implemented in software packages that can

be found in the respective papers.

The results displayed in Table 2 show that the CCT SOM outperforms the other machine learning

methods by as much as 10.83% (10.83% = 99.73% - 88.9%). We can conclude that in creating an

algorithm for the classification of walking speed given EMG inputs, a CCT SOM provides the

most accurate results, with an equal amount of computational expense as most other machine

learning pattern classification algorithms.


14

Fig. 10: Visualization of the two-dimensional topology map outputted by the SOM. This visualization does

not reflect the exact CCT SOM created in this study, but simply serves for visualization purposes.

5. CONCLUSION

The final goal of this research project is to further research on a more optimal myoelectric

prosthetic control scheme that would be able to achieve accuracy in classification tasks while

remaining simple enough to be implemented in inexpensive devices. Results from this

preliminary study suggest that self organizing maps can be used to accurately classify human

movement given EMG input data. Given their promising accuracy in walking speed

classification, low computational expense compared to networks that output spatial coordinates,

and fast training time, SOMs could potentially form the basis for a more optimal myoelectric

control scheme for prosthetics. However, this study utilized a limited data set and did not assess

every method of EMG analysis for prosthetics.

Therefore; further work should include a more comprehensive study to cover a wider range of

movements with high performance data acquisition and better statistical significance. In addition,

the possibility of testing EMG data with different neuroclustering techniques should be

considered, with the ultimate goal of creating a myoelectric control scheme that can successfully

integrate machine learning techniques to create a computationally inexpensive and accurate

prosthetic device.


15

REFERENCES [1] Gandolla, S. Ferrante, D. Baldassini, M. Cotti Cottini, C. Seneci and A. Pedrocchi, Mediterranean

Conference on Medical and Biological Engineering and Computing, 14th ed. Milano, Italy: NearLab,

Department of Electronics, Information, and BIoengineering, 2016, pp. 634-637.

[2] H. Herr, G. Whiteley and D. Childress, Cyborg Technology– Biomimetic Orthotic and Prosthetic

Technology. Bellingham, Washing-ton: SPIE Press, 2003, pp. 2-35.

[3] A. El Gehani, S. Mohamed, H. Busedra and A. Gawedar, “EMG Pattern Recognition System Based

on Self-Organizing Map for Myo-Electric Prosthesis”, ICECE, 2013.

[4] C. Pattichis, C. Schizas and L. Middleton, “Neural network models in EMG diagnosis”, IEEE

Transactions on Biomedical Engineering, vol. 42, no. 5, pp. 486-496, 1995.

[5] G. McGimpsey and T. Bradford, Limb prosthetics services and devices, 1st ed. Worcester,

Massachusetts: Bioengineering Institute Center for Neuroprosthetics, 2008.

[6] H. Jahani Fariman, S. Ahmad, M. Hamiruce Marhaban, M. Ali Jan Ghasab and P. Chappell, “Simple

and Computationally Efficient Movement Classification Approach for EMG-controlled Prosthetic

Hand: ANFIS vs. Artificial Neural Network”, Intelligent Automation & Soft Computing, vol. 21, no.

4, pp. 559-573, 2015.

[7] A. Balbinot, A. Junior and G. Favieiro, “Decoding Arm Movements by Myoelectric Signal and

Artificial Neural Networks”, ICA, vol. 04, no. 01, pp. 87-93, 2013.

[8] D. Farina, Ning Jiang, H. Rehbaum, A. Holobar, B. Graimann, H. Dietl and O. Aszmann, “The

Extraction of Neural Information from the Surface EMG for the Control of Upper-Limb Prostheses:

Emerging Avenues and Challenges”, IEEE Trans. Neural Syst. Rehabil. Eng., vol. 22, no. 4, pp. 797-

809, 2014.

[9] M. Seyedali, J. Czerniecki, D. Morgenroth and M. Hahn, “Co-contraction patterns of trans-tibial

amputee ankle and knee mus-culature during gait”, Journal of NeuroEngineering and Rehabilitation,

vol. 9, no. 1, p. 29, 2012.

[10] M. Reaz, M. Hussain and F. Mohd-Yasin, “Techniques of EMG sig-nal analysis: detection,

processing, classification and applications”, Biol. Proced. Online, vol. 8, no. 1, pp. 11-35, 2006.

[11] A. Au and R. Kirsch, “EMG-based prediction of shoulder and elbow kinematics in able-bodied and

spinal cord injured individ-uals”, IEEE Transactions on Rehabilitation Engineering, vol. 8, no. 4, pp.

471-480, 2000.

[12] C. Vaughan, B. Davis and J. O’Connor, Dynamics of human gait. Champaign, Ill.: Human Kinetics

Publishers, 1992.

[13] A. Hof, H. Elzinga, W. Grimmius and J. Halbertsma, “Detection of non-standard EMG profiles in

walking”, Gait and Posture, vol. 21, no. 2, pp. 171-177, 2005.

[14] J. Draye, D. Pavisic, G. Cheron and G. Libert, “Dynamic recurrent neural networks: a dynamical

analysis”, IEEE Trans. Syst., Man, Cybern. B, vol. 26, no. 5, pp. 692-706, 1996.

[15] G. Cheron, J. Draye, M. Bourgeios and G. Libert, “A dynamic neural network identification of

electromyography and arm trajec-tory relationship during complex movements”, IEEE Transactions

on Biomedical Engineering, vol. 43, no. 5, pp. 552-558, 1996.

[16] T. Kohonen, “The self-organizing map”, Neurocomputing, vol. 21, no. 1-3, pp. 1-6, 1998.

[17] H. Huang, Y. Liu, L. Liu and C. Wong, “EMG Classification for Prehensile Postures Using Cascaded

Architecture of Neural Networks with Self-Organizing Maps”, 2003 IEEE International Conference

on Robotics & Automation, 2003.

[18] C. Christodoulou and C. Pattichis, “Unsupervised pattern recog-nition for the classification of EMG

signals”, IEEE Transactions on Biomedical Engineering, vol. 46, no. 2, pp. 169-178, 1999.

[19] Gunther and S. Fritsch, “neuralnet: Training of Neural Networks”, The R journal, vol. 2, no. 1, pp.

30-38, 2010.


16

[20] J. Heaton, Programming neural networks with Encog3 in Java.

[21] L. Fausett, Fundamentals of neural networks. Englewood Cliffs, NJ: Prentice-Hall, 1994

[22] C. Igel and M. Hsken, “Empirical evaluation of the improved Rprop learning algorithms”,

Neurocomputing, vol. 50, pp. 105-123, 2003.

[23] A. Nicholson, “A Beginner’s Guide to Recurrent Networks and LSTMs - Deeplearning4j: Open-

source, distributed deep learn-ing for the JVM”, Deeplearning4j.org, 2016. [Online]. Available:

http://deeplearning4j.org/lstm.html. [Accessed: 01- Aug- 2016].

[24] J. Szkola, K. Pancerz and J. Warchol, “Recurrent Neural Net-works in Computer-Based Clinical

Decision Support for Laryn-gopathies: An Experimental Study”, Computational Intelligence and

Neuroscience, vol. 2011, pp. 1-8, 2011.

[25] M. Dweib, “Optimal Architecture Setting of Learning Parameters Based on Counter Back-

Propagation Neural Networks”, ARPN Journal of Science and Technology, vol. 5, no. 12, pp. 616-

619, 2015.

[26] J. Heaton, “Encog Library of Interchangeable Machine Learning Models for Java and C#”, Journal of

Machine Learning Research, vol. 16, pp. 1243-47, 2015.

[27] E. Gokgoz and A. Subasi, “Comparison of decision tree algo-rithms for EMG signal classification

using DWT”, Biomedical Signal Processing and Control, vol. 18, pp. 138-144, 2015.

OPTIMIZATION OF NEURAL NETWORK ARCHITECTURE FOR ... · OPTIMIZATION OF NEURAL NETWORK ARCHITECTURE...

Documents

Transcript of OPTIMIZATION OF NEURAL NETWORK ARCHITECTURE FOR ... · OPTIMIZATION OF NEURAL NETWORK ARCHITECTURE...